# [[The Environmental Aesthetics]] of [[Artificial Intelligence]]. 25 Feb 2025 type:: plan paper:: [[environmental aesthetics]] of ai - In our previous work, we characterised generative AI systems like Midjourney as a new type of [[artistic media]]. [[Machina naturans]]. - This characterisation allowed for one way we might think of how we might aesthetically appreciate the output of generative AI systems: - Generative AI systems present a particular form of [[dynamic recalcitrance]]; we can admire a prompter's skill/effort in grappling with such recalcitrance. - In this paper, we want to point out another facet of the [[aesthetic appreciation]] of generative AI. Specifically, we want to argue that knowledge of what generative AI systems are allows for a mode of appreciation similar to how Carlson suggests we appreciate the [[natural environment]]. 1. Carlson’s [[environmental model]] of [[aesthetic appreciation]] - Carlson: knowledge of what the [[natural environment]] in fact is (use the quote) feeds into [[aesthetic appreciation]]. - His characterisation of the environment: forces, systems, effects, - these things can be understood through science - but we might also think that someone can develop a practical understanding of these same sorts of things: gardeners, farmers, etc. - while Carlson focuses on contemplation, some of the stuff he says about [[common sense]] knowledge is suggestive. 1. What Generative AI Systems in [[fact are]]. - Start with Amodei quote, maybe throw the other one from Twitter yesterday that I just saw. - This quote is evocative, but at the same time, when we think about how these systems behave, we can see some parallels with nature. - Do the usual natura, machina stuff. - Whereas [[scientific knowledge]] in the environmental case is biological or ecological knowledge (etc.), scientific knowledge in this case would be about the architecture of generative AI. - [Add some detail about how models are trained] - Talk about weighting, entanglement, manifold. 1. Appreciation for generative environment/systems as what they in fact are. - Looking at an image generated by Midjourney. - It looks like an image made by a human and suggests that it should be appreciated as such. - We suggest an alternative: they should be looked at as the outputs of a latent, entangled system in the way we have just characterised. - Something about an understanding of art history and techniques as feeding into this. Being able to see elements from different periods in the same piece, - being able to understand clusters or entanglement. - Understanding that it is more difficult to get Midjourney to spit out some sorts of styles, or mixes of styles than others. 1. Objections. - Doesn’t this entail that only computer scientists will be able to truly appreciate AI art? - Two reasons to think not: - It might just be a matter of public understanding of new technology being behind the times. These sorts of systems are relatively new; perhaps once we understand the basics of generative systems more widely, in the same way that biological knowledge or ecological knowledge are now widespread, but were at some point not. - It might be that we are in the infancy as to what these things can do. - A more substantial riposte: through using these systems, we gain a practical knowledge of generative systems, similar to how a gardener might with the natural environment. - How Images Are Generated: Putting it together, when you prompt a diffusion-based system: It encodes the prompt into an embedding (capturing what you asked for, in a broad sense). It initializes a latent array filled with random noise. This is the canvas. The diffusion U-Net, guided by the prompt embedding, repeatedly denoises the latent. Over dozens of steps, recognizable structures emerge from the noise, increasingly resembling a coherent image. The model “knows” how to do this because it was trained on countless examples of noised images and their denoised counterparts. Once the latent is sufficiently refined, the decoder transforms it into a final image. This process can be likened to a master painter starting with an abstract underpainting and adding details iteratively. However, unlike a human, the diffusion model’s steps are learned from data and follow a probability-driven pattern rather than a conscious creative plan. This can lead to averaging effects: the model often gravitates toward producing the most statistically likely version of an image given the prompt. If many training images of “an apple on a tree” showed a generic red apple on a brown branch against green leaves, the generated result may default to that kind of composition and style, unless the prompt specifies otherwise. In other words, the architecture is biased toward recreating the distribution mean of the training data for a given prompt. This is one reason AI images can feel uninspired or expected – they are excellent pastiches of the training examples, but rarely venture outside of them by default.In summary, diffusion model architectures, by design, produce images that reflect the patterns and limitations of their training. The use of a latent space can smooth out unique details, the iterative noise removal process encourages consistency (sometimes at the cost of creativity), and text-conditioning guides the image toward common interpretations of the prompt. Next, we discuss the training process in more detail and how it contributes to the final aesthetic. # Plan to Paper Prompt ## Prompt: Plan to Paper PROMPT BEGINS: 1. Context and Purpose You have two source documents: Document A: A plan for an analytic philosophy paper. Document B: A draft and/or collection of notes, which may be lengthy and disorganised. Your task is to write a new analytic philosophy paper that follows the structure and goals outlined in Document A, while reusing as much text from Document B as possible. If some parts of Document A are not covered by any relevant text in Document B, you may compose new text. However, always prioritise splicing together or repurposing text from Document B over writing entirely new material. The final paper should be approximately 3000 words in length. Requirements Adhere to the plan: The new paper must reflect the points and structure detailed in Document A. Maximise reuse of text from Document B: Incorporate sentences, paragraphs, phrases, or clauses from Document B to meet the requirements of the plan. Minimal new text: Provide original wording only when there is insufficient or no relevant text in Document B to address a particular aspect of the plan. Maintain a consistent style: Ensure the final paper is written in an analytic, formal tone, consistent with typical academic writing. Avoid overtly dramatic or poetic language. Length: The paper should be approximately 3000 words. Output format: Provide two versions of your final output: A continuous text of the academic paper. The same text rendered as bullet-pointed sentences, allowing easy visualisation of the structure. Process and Reasoning Approach the task systematically: First, review Document A (the plan) to determine the required sections and main points. Then, review Document B (the draft/notes) to identify text suitable for these sections. Integrate the reused text into a coherent paper. Where necessary, revise wording minimally for flow, grammatical accuracy, and consistency. If a segment of the plan is not addressed in Document B, compose new text in an analytic, neutral tone. Ensure no substantial repetition unless needed for coherence. Use headings or sectional divisions if appropriate. Following completion of the 3000-word paper, create a bullet-pointed format where each sentence is a separate bullet point, preserving the same sequence as in the continuous text. Stylistic Considerations Use British English spelling and conventions. Employ clear, direct, and academically neutral language. Avoid speculative or value-laden terms such as “posits” or “inherent”, as well as overly emphatic words like “critical” or “pivotal”. Maintain a consistent level of formality suitable for an academic audience. Output Specification Part One: A coherent academic paper of approximately 3000 words, following the structure from Document A and utilising text from Document B wherever possible. Part Two: The same text reformatted into a bullet-point list, with each sentence as a separate bullet point, presented in the same order. Instructions to the Model Read and integrate the text from both documents carefully. Ensure that the final paper’s structure and content remain faithful to Document A. Copy, splice, and adapt text from Document B before composing new passages. Insert original passages only when necessary to fulfil Document A's requirements that are absent in Document B. Maintain an objective, analytic tone throughout. Final Check Confirm that the final paper is cohesive, logically organised, and addresses all points in Document A. Verify that each statement has been drawn from Document B whenever feasible, with minimal new wording. Confirm that two versions (continuous text and bullet-pointed sentences) are provided. Finally, provide a report on precisely what you have done. /END OR PROMPT ## Prompt: Plan Enhancer Context and Purpose You have two source documents: Document A: A plan for an academic paper. Document B: A draft and/or collection of notes, which may be lengthy and disorganised. Your task is to create a more detailed plan that reflects the overall structure of Document A while integrating relevant content from Document B. Identification of Topics from Document B Examine Document B carefully, identifying all potentially relevant topics, arguments, analogies, quotations, examples, or other useful points. Compile these findings into a detailed list of potential topics. Assign a unique letter (e.g., A, B, C, ...) to each topic so that the user can easily refer to them. Each topic entry should briefly summarise the idea or argument and, if possible, reference the corresponding text in Document B (e.g., a key sentence, phrase, or section). Reconstructing the Existing Plan (Document A) Recreate the plan’s structure (chapters, sections, subsections) as it appears in Document A, preserving its headings and order. Present each section or subsection clearly, using a numbered or otherwise systematic format. Proposed Integration of Topics After listing the topics and recreating the plan, ask the user which of the identified topics (from the lettered list) they would like to include under each section or subsection of the plan. The user will then indicate which topics belong in which sections. Updating the Plan Based on the user’s choices, update the plan to show which topics fit under each section or subsection. Strive to incorporate the user’s selected topics into the plan in a coherent and concise manner. Incorporate only minimal wording changes to reflect logical connections between plan sections and the selected topics. Do not refer to the ideas added by the letters they have been assigned. the sole purpose of the letters was for easy reference by the user. Always write out the entire plan, even for sections to which no changes have been made. Stylistic and Formal Requirements Use British English conventions. Maintain a clear, direct, and analytically neutral tone. Avoid overly emphatic or value-laden expressions (e.g., “critical,” “inherent,” “pivotal”). Ensure that all newly composed text is kept to a minimum, and focus on reusing, clarifying, or reorganising the text from Document B. Instructions to the Model Identify topics from Document B comprehensively. Label each identified topic with a unique letter and provide a concise description. Reconstruct the plan from Document A without alterations to its fundamental structure. Invite user input on where each lettered topic should go in the plan. Incorporate the user-selected topics into the plan coherently and succinctly, using or minimally modifying the text from Document B. Present the updated plan in a way that clearly shows how the selected topics have been integrated. Final Output Your final output should include: The List of Topics: A comprehensive list of all identified topics from Document B, each labelled with a unique letter and briefly summarised. The Reconstructed Plan: A reiteration of Document A’s structure . An Invitation for User Input: A prompt asking the user which topics they would like to include in each section or subsection. (After receiving user input) The Updated Plan: A revised version of the plan that integrates the chosen topics under their respective sections. This new version will be written out IN FULL, even for sections which have not been added to. # Draft Plan paper:: environmental aesthetics of ai type:: plan ## Introduction - We propose that there is an instructive parallel between the way we come to appreciate the natural environment and the way we might come to appreciate generative AI. - Carlson highlights an approach to nature grounded in understanding what nature is and how it works, we suggest a similar approach for AI outputs, where we consider the conditions and forces shaping them. --- ## 1. The failure of function as a means of aesthetically appreciating generative AI - Function-based theories, such as those that locate aesthetic value in how effectively an object fulfils a defined purpose, run into problems when applied to generative AI. - These systems do not have one clear function. They might produce text, images, or serve multiple roles (from tutoring to code generation). - It is therefore difficult to evaluate their aesthetic qualities solely through whether they meet a singular purpose. - Some accounts, grounded in how an object’s form matches its intended function, also encounter difficulties in the open-ended context of AI, where the scope of possible uses is constantly expanding. --- ## 2. The Natural Environmental Model - Carlson argues that knowledge of what the natural environment in fact is feeds into aesthetic appreciation. - He emphasises appreciating environments as dynamic, shaped by forces and systems, which can be understood scientifically or through everyday familiarity. - We may observe a similar principle in the context of generative AI: appreciating its outputs involves recognising them as the products of an evolving, partly autonomous system. - A brief hint at broader knowledge: just as scientific knowledge of nature can be complemented by a more common-sense, practical grasp, so too can one gain insight into AI through hands-on, non-specialist experience. - Spinoza’s notion of a self-generating environment (natura naturans) sheds light on the idea that something can be a continuously productive force rather than just a static object. We draw a preliminary parallel here to later sections, suggesting that generative AI may similarly be viewed as an ongoing process rather than a straightforward tool. --- ## 3. What Generative Systems are - One way to see how these systems behave as evolving environments is through remarks likening their development to a growing process rather than traditional programming. - For instance, it has been suggested that large models are “grown,” using training objectives as a kind of guiding “light,” so that they develop emergent properties much like organic entities do. - Thinking of generative AI as a new form of “machina naturans” highlights its capacity for ongoing creation. The system’s outputs are not strictly pre-specified but arise from shifting interactions among data and architecture. - Internally, these systems rely on latent spaces or manifold representations. The model “learns” a sprawling, multidimensional structure from which it draws its generated responses. - By analogy, the dataset and learned patterns act like a “soil”: if certain styles or subjects appear abundantly in training, they become easier for the model to produce. - This manifold perspective can also help explain why some prompts or styles are more difficult to achieve, much like certain plants thrive only under precise soil conditions. - Related to this soil analogy is the idea of adjusting overall parameters—akin to balancing pH—when emphasising or suppressing particular features in an output. - Indeed, once trained, the model can be navigated (via prompts) like a digital environment whose internal “terrain” emerges from how the data were structured. This environment can be surprisingly vast and varied. --- ## 4. Appreciating Generative systems - When confronted with, say, an image generated by Midjourney, we could regard it as a human-made artwork. However, an alternative stance is to see it as the fruit of an evolving, partly self-directed process. - In some ways, it recalls how certain carefully landscaped places blend nature and design. One might compare it to a form of gardening, yet here the “garden” is an environment made up of algorithms, data, and user inputs. - Extending that analogy, different cultural or aesthetic elements in the training data may combine in unexpected ways, akin to a cultivated yet semi-wild landscape. --- ## 5. Objections - A potential objection is that only those with advanced technical expertise can appreciate generative AI in the manner described. - Over time, however, public familiarity with AI may increase, similar to how basic ecological knowledge has become more widespread. - Hands-on experimentation can equip lay users with a practical sense of a system’s workings. This recalls the way a gardener gains knowledge of soil, weather, and plants without necessarily having a formal science background. - Another worry is that if such technologies continue evolving, new aspects might complicate aesthetic appreciation. Yet this evolution might equally deepen public discourse, suggesting that our understanding of AI’s capabilities and constraints will keep expanding. - Engaging with AI through repeated prompts, trial and error, and a growing cultural conversation means that aesthetic appreciation does not require specialist credentials. Rather, it can flow from practical experience and a willingness to view these systems as active, dynamic environments. # message to elsa Ciao, I have read your paper through a couple of times now, and to be honest I don't have very much to say at all right now. Philosophically it is very good, and it is extremely well written. I am going to take a break for a couple of hours now and will either look at this again later today or tomorrow morning. Is that ok? Any comments you get will be on the small details side rather than big sweeping changes. Before I forget though, there's a missing 'be' after the 'will' in this sentence: > Another example that is worth examining is Fountain (1917) by Marcel Duchamp, the pioneering artwork that initiated the “ready-mades” kind, ⁠⁠bearing in mind the ontological complexity of this artwork (Evnine 2013) that will not an immediate concern for this paper.⁠⁠ > Btw, my plan is to be in the office sometime around 11am tomorrow, if you're around between then and 14.30, we can talk about this face to face. I can just as well do it via email though, so whatever suits you.